Mojo function

generic_fused_qk_rope_bshd_continuous_batch

generic_fused_qk_rope_bshd_continuous_batch[dtype: DType, //, *, interleaved: Bool, target: StringSlice[StaticConstantOrigin]](q_proj: TileTensor[dtype, address_space=q_proj.address_space, linear_idx_type=q_proj.linear_idx_type, element_size=q_proj.element_size], kv_collection: ContinuousBatchingKVCacheCollection, freqs_cis: TileTensor[dtype, address_space=freqs_cis.address_space, linear_idx_type=freqs_cis.linear_idx_type, element_size=freqs_cis.element_size], layer_idx: UInt32, valid_lengths: TileTensor[DType.uint32, address_space=valid_lengths.address_space, linear_idx_type=valid_lengths.linear_idx_type, element_size=valid_lengths.element_size], output: TileTensor[dtype, address_space=output.address_space, linear_idx_type=output.linear_idx_type, element_size=output.element_size], context: DeviceContextPtr = DeviceContextPtr())

Performs a fused RoPE projection for Q and K projections.

We have a manually fused QKV projection with mo.opaque dtypes in our Llama model. Due to a limitation in custom op definitions, we can't declare both a tensor and opaque dtype as output from a custom kernel. This requires us to only note Q_proj as an output from the QKV projection. If we immediately follow the QKV proj kernel with a RoPE kernel applied to K, we'll get a race condition because the graph compiler doesn't know about the dependency between these kernels in the graph definition. Here we fuse the RoPE kernel applied to Q_proj with K_proj, so K_proj RoPE is only executed after QKV completes.

Args:

q_proj (TileTensor[dtype, address_space=q_proj.address_space, linear_idx_type=q_proj.linear_idx_type, element_size=q_proj.element_size]): Query projection tensor of shape [batch, seq_len, n_heads, head_dim].
kv_collection (ContinuousBatchingKVCacheCollection): The continuous batching KV cache collection.
freqs_cis (TileTensor[dtype, address_space=freqs_cis.address_space, linear_idx_type=freqs_cis.linear_idx_type, element_size=freqs_cis.element_size]): Frequency tensor for RoPE of shape [max_seq_len, head_dim].
layer_idx (UInt32): The layer index for accessing the correct cache.
valid_lengths (TileTensor[DType.uint32, address_space=valid_lengths.address_space, linear_idx_type=valid_lengths.linear_idx_type, element_size=valid_lengths.element_size]): Tensor of shape [batch] containing the valid length for each sequence. RoPE is only applied to positions within these lengths.
output (TileTensor[dtype, address_space=output.address_space, linear_idx_type=output.linear_idx_type, element_size=output.element_size]): Output tensor for Q with RoPE applied, same shape as q_proj.
context (DeviceContextPtr): Device context pointer for execution.